How to use Claude for ads: skills, MCP and a human on the button

We run ad production for SmartSolar with Claude, in Claude Code and with a set of written skills. This page covers how: what skills and MCP each do, what we connect, what we keep out of reach, and where a person stays on the button. It is our setup, described plainly. Definitions come from Anthropic’s own documentation, which we link.

A skill is the written method for a job, while MCP connects the AI to tools and data. Keeping those roles separate makes it easier to review what a request can actually do. The detailed skills and MCP comparison explains where instructions end, where a tool call begins and when a subagent is useful. These components are building blocks for an ad workflow, rather than evidence that a whole production system is available.

How a skill loads

1. 01Name and descriptionAlways in context
2. 02The SKILL.md instructionsWhen the task calls for the skill
3. 03Files and scriptsOnly when the instructions point to them

What we write as skills

One skill per job: the ad factory method, the editor and renderer, the voice pipeline, avatar stills, skeleton ads and synthetic persona tests. A skill is written method, sharpened ad by ad. We keep the engine and the market apart. The engine is the universal method. Anything specific to a market, such as numbers, claims, voices and casting, lives in a market profile that collects its own lessons. We never write one market into the engine.

Today these skills run in our own studio. Our skill pack in your own Claude Early access is our method set up in your own Claude. Early access: we set it up with you. Coming: we’re building it. The connected version, The factory in your AI Coming, is the plan to offer our skills, your library and review in your own AI over MCP, starting with Claude.

How to connect the models to Claude safely

The models we use for video, stills, voice, music and speech-to-text are separate services. We reach them through scripts and connectors, and a few rules keep that safe.

Keys live in a key folder, scripts read them, and they are never printed, logged or pasted into chat. If a key has to be entered, the person types it locally. For the provider-specific setup, connecting ElevenLabs to Claude separates the official server routes and explains why the available tools need inspection before access is granted.

Generation is asynchronous. An agent session can end while a provider job continues, so we submit a job and poll it, with a state file. A job that was sent or has succeeded is never sent again, because that bills twice. A killed call doesn’t kill the job at the provider, so we check status by request id before we retry. And we fetch outputs right away, because output links are temporary.

Every job has a hard dollar cap, and its id and cost are logged the moment it is sent. We estimate first, make one test clip, get it approved and only then run the batch. The model-specific rules are in how to use Seedance for ads, ElevenLabs for ad voiceovers and Nano Banana for ads. Models in Claude Coming is the plan for reaching video, stills and voice through one Turntake connection, with Cost shown before you run Coming showing the cost in credits first and Bring your own tools Coming for the accounts you already have. The Turntake models page describes that plan, and Your AI shows the connection.

One generation job

1. 01Estimate
2. 02One test clip
3. 03ApproveA person says yesYou
4. 04Run the batch
5. 05Log the job id and costThe moment it is sent
6. 06Poll with a state fileNever resend a sent job
7. 07Fetch the outputLinks are temporary

In our production in October 2026, each session reads a small index, current status and the relevant task. Decisions and their evidence go into files before the session ends. The project memory guide owns the file structure, and briefing several agents covers bounded parallel work. A handoff should make the next action clear without requiring the next person to replay the conversation.

In our own studio in October 2026, Claude revises a media recipe and a renderer makes the next take. The Claude video editing guide explains that boundary, while video feedback notes keep the requested change attached to a specific moment.

In our production in October 2026, account analysis is read-only, a person runs the upload and everything is created paused. The Meta analysis boundary explains why a recommendation and permission to change an account are separate decisions.

Claude mayA person does

• Claude mayRead ad accounts and drivesA person doesRun the upload
• Claude mayAnalyze and reportA person doesSwitch an ad on
• Claude mayArchive, when you say deleteA person doesAnything that costs money or changes an account

Choose the bounded production task before choosing which tools can act on it.

A guardrails checklist

• A hard cap per agent, and a cost log with the job id written immediately.
• At most five agents in parallel, and a disk budget for temporary files.
• Agents write only in their own folders. Shared scripts are copied, not edited, while parallel agents run. One round, one desk.
• Keys stay in the key folder, and only scripts read them.
• The ad account is read-only for agents, a person runs the upload, and everything is created paused.
• A checker that doesn’t fix. A separate agent verifies and never repairs. Numbers in an agent’s report are counted by a script over the agent’s files, because one agent wrote that a cap was reached after seeing only a fraction. A critic recomputes the main numbers.

With those rules in place, Claude can carry the chain described in how to make AI UGC ads. The rules are what keep it from carrying the budget too.

What should Claude own in an ad round?

Start with the next concrete decision, rather than asking for an entire campaign. A script brief, an approved voice sample and a revision request have different inputs and different definitions of done. In our production in October 2026, the task file identifies the output and the boundary before an agent starts. That makes a request reviewable even when the implementation uses several tools.

Source: our own production rules, October 2026.

The unit matters because the task may fail between stages. A good script does not establish that a model will accept the planned reference. A successful model job does not establish that the downloaded file contains the required shot. A rendered file does not establish that the graphic and music survived the edit. Keep the acceptance evidence attached to the actual output, so the next stage can check it without guessing.

Separate creative judgment from mechanical work. Choosing the message, interpreting research and deciding whether an animation communicates the right idea need review. File inventory, duration checks and matching an output to a requested path are narrower jobs. Written criteria make those jobs suitable for scripts or bounded agents, while keeping the creative decision with the person responsible for the round.

An incomplete handoff should say what remains uncertain. If a source could not be read, the fact is unconfirmed. If an image is approved but motion is not, that is a still approval. If a reviewer accepts the voice but has not inspected captions, the final take still needs review. These distinctions prevent a green status at one step from spreading into an unsupported claim about the whole ad.

The practical benefit of using Claude across the chain is continuity of instructions and evidence. Each stage can start from a small, explicit brief, then return the file and the checks that the next stage needs. The final decision remains attached to the final take rather than to a confident message in a chat.

Which Claude production guides should you use next?

Use the detailed guide for the task in front of you. Each owns one question, so you can work through a specific production decision and return to this overview for the wider workflow.

• Claude video editing
• ElevenLabs in Claude
• Read-only ad analysis
• Skills vs MCP
• Seedance prompt skills
• Stills in Claude
• Parallel agent briefs
• Independent output review
• Project memory files
• Edit recipe contracts
• Approved version history
• Render queue states
• Machine validation gates
• Generation retry records

An ad round also needs a clear point where creative exploration stops. Once a script or shot is approved, a revision brief should identify the change and the dependencies it affects. A voice change may require timing review; a wording change may require claim review. Treat that dependency check as part of planning the next take, rather than discovering it after the render.

Who wrote this We made the ads first, for SmartSolar in Denmark, Sweden and Ireland. Then we built Turntake. Where Turntake comes from